A Conceptual Blending Approach to the Generation of Cognitive Scripts for Interactive Narrative

Intelligent Narrative Technologies: Papers from the 2013 AIIDE Workshop (WS-13-21)

Justin Permar and Brian Magerko {jpermar, magerko}@gatech.edu Georgia Institute of Technology 225 North Ave NW Atlanta, GA 30332 USA

Abstract prehension, here we focus on an explicitly temporal representation of activities. Our method of modeling a process This paper presents a computational approach to the that produces script blends from input scripts is based on the generation of cognitive scripts employed in freeform theory of conceptual blending. activities such as pretend play. Pretend play activities involve a high degree of improvisational narrative con- Conceptual blending (also known as “conceptual integra- struction using cognitive scripts acquired from everyday tion theory”) has been proposed as a basic cognitive op- experience, cultural experiences, and previous play ex- eration underlying our capacity for creativity as well as periences. Our computational model of cognitive script mundane meaning-making (Fauconnier and Turner 2003). generation, based upon conceptual integration theory, Pretend play is a narrative domain where numerous cre- applies operations to familiar scripts to generate new ative feats regularly appear, including the use of real ob- blended scripts. jects as pretend objects, diegetic and extra-diegetic context switches, and script blending, such as playing cars and 1 Introduction Godzilla wrecking the city. However, script blending is cer- tainly not confined to pretend play, but instead occurs perva- People regularly engage in social activities that conform to sively throughout creative works. Music composition, film cognitive scripts, which are temporally ordered sequences pastiche and narrative sub-genres are all examples of art of actions and events that represent a narrative experience works created from multiple sources. (Schank and Abelson 1977). Our everyday experiences of- Conceptual blending is a theoretical abstraction encom- ten involve modifications to cognitive scripts, where features passing a group of related cognitive processes, including from multiple scripts are meshed together to help us make analogical reasoning, mental modeling, and similarity (Fau- sense of the world or create an interesting new narrative ex- connier and Turner 1998). Fauconnier and Turner’s net- perience. For example, children pretend play as heroes or work model of conceptual blending involves four concep- villains based upon situations seen in superhero films. Al- tual spaces: a generic space, two inputs spaces, and a blend though this narrative process is ubiquitous, formally under- (output) space. In addition to the overall structure of these standing how to combine narrative scripts to elicit new, cre- spaces, conceptual blending also proposes standard pro- ative scripts is an unsolved problem (Magerko et al. 2009). cesses that occur for the purposes of “on-line, dynamical By better understanding how to accomplish this, we can in- cognitive work” (Fauconnier and Turner 1998). This net- form approaches to computational creativity that deal with work model is a generalization of many cognitive processes narrative understanding and generation. that produce an output derived from two inputs, such as a Our work addresses the generation of unique scripts given process of script generation using two familiar input scripts. two input scripts from an agent’s background knowledge in Conceptual blending processes that are relevant to script the narrative domain of pretend play. Playing pretend, as a generation include 1) the cross-space mapping of counter- creative act, frequently necessitates the generation of scripts part connections between input spaces and 2) selective pro- that are blended from cultural references, past play experi- jection from the input spaces to the blend space. A key result ences, and common cultural scripts (Moran, John-Steiner, of blending is that the composition of elements (in the blend and Sawyer 2003). In this paper, we present an algorithm space) from the input spaces results in relations not present that generates new “script blends”, represented as scripts. in either input space. This capability can be used, for example, by an AI agent to execute scripts that are not defined a priori. Similarly, To date, research related to conceptual blending, which an agent could use this capability to aid script recognition we review in section 2, has mostly focused on metaphor and during pretend play activities. Although scripts consist of analogy. Some research has successfully utilized concep- causal and temporal links, both of which are critical to com- tual blending for other purposes, including interactive poetry generation with Harrell’s GRIOT system and the work Copyright c 2013, Association for the Advancement of Artificial by Li et al. on the generation of fantastical and imaginary Intelligence (www.aaai.org). All rights reserved. pretend play objects (Harrell 2005; Li et al. 2012). How-

44 ever, no work has yet applied conceptual blending to cogni- that treats mappings as exclusively one-to-one. Our algo- tive structures that explicitly represent temporality, such as rithm mirrors later work that only uses one-to-one mappings. scripts. Our primary contribution is an initial algorithm that modifies two familiar scripts to produce a blended script, a Computational Models of Generative Processes capability that can inform generalized approaches to com- Holyoak and Thagard built ACME to produce analogies, putational generation of creative narratives. which at its core is a problem of determining suitable cor- This paper is outlined as follows. In section 2, we re- respondences between two inputs (1989). Holyoak and view related work in conceptual blending and computational Thagard describe categories of constraints that must be re- models of generative processes. In section 3, we present our spected in order to derive analogies, including 1) structural, script generation algorithm alongside a motivating example 2) semantic, and 3) pragmatic constraints. Structural con- in order to ground the remainder of the paper. In section 4, straints specify that an isomorphism between inputs is valid we present two additional examples illustrating details of our if, and only if, the mapping is one-to-one for any objects algorithm and approach. Lastly, in section 5 we conclude and relations (see (Holyoak and Thagard 1989) for formal with a critical discussion of our current approach, highlight definitions). They use a matching process that can estab- limitations of our algorithm, and present future work. lish mappings between relations when only the number of arguments matches, regardless of order. Our approach uses 2 Related Work a matching process that uses both the number and order of Conceptual Blending arguments, but slightly relaxes the requirements from exact matches for all arguments to exact matches on all arguments Zook et al. propose a formal model of pretend object play except one. based on a two-space model of conceptual blending (Zook, Falkenhainer et al. developed SME as an implementation Magerko, and Riedl 2011). The model contains a concep- of analogical reasoning based upon Gentner’s structure map- tual space comprised of real objects in a real domain and ping theory (Falkenhainer, Forbus, and Gentner 1989). SME a second conceptual space comprised of imaginary objects is an implementation of analogy based on structural similar- in a pretend domain. They propose a process in which the ity between two representations (of situations). Structural last step blends attributes of real and pretend objects to pro- similarity is achieved by aligning elements of two (simi- duce a blended object, such as a pretend sword. Both our lar) representations using two primary constraints: 1) one- work and the work by Zook et al. employ a process involv- to-one mapping, and 2) parallel connectivity (retaining the ing substitution operations during blending, but their work order of arguments for a predicate). SME, as pointed out does not involve inputs with a temporal dimension, such as by (Holyoak and Thagard 1989) requires an exact match of scripts, which necessitates a significantly different approach predicates (part of the predicate calculus representation fa- to counterpart mapping. vored by Forbus et al.) during mapping. Closely related to Veale et al. provide a computational model of concep- SME is MAC/FAC, where the MAC stage uses a computa- tual blending called Sapper that uses a semantic network to tionally inexpensive non-structural filter to reduce compu- model concepts and relations between concepts and a pro- tational complexity (Forbus, Gentner, and Law 1995). Our cess of spreading activation to explore the conceptual do- approach is closely related to SME, as we use the notions of main (Veale, O’Donoghue, and Keane 2000). Their goal predicate matching and parallel connectivity to assert equiv- during counterpart mapping is to establish a sub-graph iso- alence. A key difference is our use of path representations morphism between the two input spaces, which allows “pro- during structural alignment. jection” of mapped properties between objects. Veale et al. base the means for establishing the sub-graph isomorphism on a structural mapping that is derived from Gentner’s 3 Script Blending Algorithm structure-mapping theory of analogy and the property of sys- In order to guide this section, we present an example blend tematicity, which is the idea that particular (analogical) map- between two scripts familiar to Western culture: a stage- pings are defined by the existence of higher-order relations coach ride script, shown in Figure 1, and a guided tour bus (Gentner 1983). Sapper takes as input a semantic network, tour script, shown in Figure 2. Our script representation is which does not represent the temporal nature of scripts that a directed acyclic graph (DAG), a restriction of a general is central to our problem and approach. graph structure that we require in order to establish a finite Goguen and Harrell developed the ALLOY blending en- number of “paths” through a script. The script is read top- gine, noting that concepts (and relations between concepts) down, with an initial event beginning the script. The edges are a representation which does not fully support describing between nodes are directed, indicating a temporal ordering “the structure of things” (Goguen and Harrell 2004). Thus, among actions in the script. The predicates should be inter- they focus on a variant of conceptual blending that they call preted as actions. Parentheses surround the argument list for structural blending or structural integration. Semiotics, an the predicate, henceforth referred to as the entity list. All en- integral part of the ALLOY blending engine, includes the tities are capitalized in order to visually distinguish between notion of morphisms that map between signs in a source entities and predicates. The last noteworthy feature illus- space and signs in a target space. ALLOY has a notion of trated in Figure 1 is the inclusion of a floating-point number partial mapping between elements in the source and target (0.6) just after the Stagecoach entity in the “enter” predicate. spaces, which is a significant differentiation from later work This number is an iconicity value, which is a real number in

45 1 arrive(Driver, Boarding Location) tains the stagecoach script as the source input and the tour bus tour script as the target input. In our algorithm, the target script is always the input that is modiﬁed in order to generate wait(Driver) 2 3 enter(Passenger, Stagecoach (0.6)) a script blend (an approach inspired by metaphor).

return( Passenger, tell(Passenger, Driver, White 4 5 Counterpart-mapping Phase Stagecoach) House) The first algorithm phase, known as counterpart-mapping, identifies all paths through the input scripts, given a start departs(Driver, Passenger) 6 node and an end node in each script. Paths are identified using a modified depth-first search that continues to search for arrival notice( Driver, 8 7 converse(Driver, Passenger) Passenger) paths until all edges in the graph have been traversed. This depth-first search step is guaranteed complete due to the use

arrive(Driver, Passenger, White of DAGs. We attempt to model the notion of “similar activi- 9 House) ties” as a formalism of structural similarity between a pair of paths in the two input scripts, which retains the temporal or- Figure 1: Stagecoach source input script dering of nodes as a form of “higher-order” structure when comparing nodes during the next step.

arrive(Passenger, Boarding Loca- After all paths have been found, the algorithm establishes 1 tion) counterpart mappings via three node match rules: 1. Exact-match: the predicate in the source and target nodes 2 tell(Guide, Passenger, Board Bus) match (using literal equality) and both the order and number of entities matches. Note that this rule is a require-

3 board(Passenger, Tour Bus) ment for both predicate equality and parallel connectivity (from structure-mapping theory). Note that nodes common to both scripts will be mapped, which provides for tell(Guide, Passenger, 4 Welcome) temporal coherence. 2. Predicate-match: a relaxation of the exact-match rule 5 depart(Guide, Passenger) where exactly one entity differs. The strictness of exactly one different entity is an effort at ensuring coherence of generated blends. For example, eat(Dog, Dog 6 visit(Guide, Passenger, Landmark) Food) does not have a predicate-match with eat(Person, Chinese Food) because two arguments differ. However, question( Passenger, arrive(Guide, Passen- Guide, What is the 7 8 ger, Boarding Loca- eat(Dog, Chinese Food) would match with eat(Person, next stop?) tion) Chinese Food) because exactly one argument differs. 3. Ordered-entity-match: similar to the exact-match rule, but Figure 2: Tour bus target input script predicates are allowed to differ. That is, the requirement is for corresponding arguments to match exactly. the range [0,1] that indicates the uniqueness of an entity (or Holyoak and Thagard built ACME as an analogical rea- predicate) in a script within the context of a script knowl- soning system. Our approach differs primarily by requiring edgebase (Magerko, Dohogne, and DeLeon 2011). For ex- what Falkenhainer et al. term “parallel connectivity”, with ample, a stagecoach (in the modern Western world) is com- a slight relaxation from exact matches for all arguments to monly used in the context of a horse-drawn trip within a par- exact matches on all arguments except one (in the case of ticular locality, and thus is assigned an iconicity value of 0.6. the ordered-entity-match rule). Predicates can also be iconic; for example, a “pitch” action is During counterpart-collection, every pair of paths (taking nearly exclusively associated with baseball, and thus would one from the source and one from the target to make a pair) be assigned a high iconicity value in a baseball script1. is used for pairwise comparison of nodes. This comparison Our algorithm contains three phases: 1) counterpart map- technique enforces temporal ordering, a feature of scripts ping, 2) mapping selection, and 3) mapping application. At that must be retained for comprehension. Each node in the a high level, our approach combines the path representa- source and target is tested for a match using the three rules tion mentioned by Veale et al. with the notions of one- above. If a match is found between a node ni in the source to-one mapping and parallel connectivity from structure- input and node nj in the target input, a candidate counterpart mapping theory (Gentner 1983; Veale, O’Donoghue, and mapping is proposed. An intentional restriction our algo- Keane 2000). As mentioned above, the blend example con- rithm places upon further matches is that matches are only possible using nodes that are successors of nodes ni (e.g., 1“Pitch” would also have a high iconicity value in a script for ni+1,...) and nj (e.g., nj+1,...). That is, given a mapped the game of Cricket. This demonstrates the importance of context pair of nodes (nx, ny), the only potential additional matches of a knowledgebase when calculating iconicity values. involve nodes in the source in the set {nx+1, . . . , nS} and

46 arrive(Driver, arrive( Passen- 1 arrive(Driver, Boarding Location) Boarding Lo- 1 1 ger, Boarding cation) Location)

enter( Passenger, tell( Guide, Passenger, 2 tell(Guide, Driver, Board Bus) wait(Driver) 2 3 2 Stagecoach (0.6)) Board Bus)

return( tell( Passenger, board( Passenger, Tour 3 board(Driver, Tour Bus) Passenger, 4 5 Driver, White 3 Bus) Stagecoach) House)

tell(Guide, Driver, departs( Driver, tell( Guide, Passenger, 4 6 4 Welcome) Passenger) Welcome)

arrival notice( converse( depart( Guide, Passen- 5 depart(Guide, Driver) Driver, Passen- 8 7 Driver, Passen- 5 ger) ger) ger) arrive( Driver, visit( Guide, Passen- 6 visit(Guide, Driver, Landmark) 9 6 Passenger, ger, Landmark) White House) arrive( Guide, question( Driver, question( Passen- arrive(Guide, Driver, Passenger, Guide, What is the 7 8 ger, Guide, What 7 8 Boarding Location) Boarding Lo- next stop?) is the next stop?) cation) Figure 4: “Stagecoach” + “Tour bus” output blend Figure 3: Stagecoach and Tour Bus Counterpart Mappings

nodes in the target in the set {ny+1, . . . , nT } where S is the sets in order of descending set size. For example, if the number of nodes in the source path in the current path pair counterpart-mapping phase found 5 mappings for path and T is the number of nodes in the target path in the current pair (pathS1, pathT 1) and 2 mappings for the path pair path pair. In this manner, any nodes that match under the (pathS2, pathT 2), then these two sets of mappings are constraint of temporality are found for each path. Similar to sorted in descending order (a set of 5 pairs then a set of 2 the approach by Zook et al., we identify attributes common pairs). Mappings are used so long as they do not specify to both inputs and employ a process of attribute substitution previously-mapped nodes, thereby ensuring a one-to-one during blending. A key difference, however, is the temporal node mapping. nature of scripts, which requires a significantly different approach to counterpart mapping. An example pairwise path As mentioned above, our current mapping selection heuris- mapping comparison is show in Figure 3. The “active” path tic is to choose mappings in order of descending set size, through each script is shown in bold. which is a greedy algorithm. Our motivation for using a Figure 3 can be used to illustrate the restriction of tempo- greedy selection algorithm is to have a large set of selected rality used during counterpart mapping. The “arrive” nodes mappings at the end of the mapping-selection phase. This are mapped to each other. Consequently, the only nodes may prove not to be the best heuristic for mapping selection, available for further counterpart mapping match tests are which is an area for future work. nodes after the “arrive” nodes. Mapping-selection Phase Modify-target Phase Mapping-selection takes as input the pairs of counterpart mappings from the counterpart-mapping phase in order to The modify-target phase uses selected mappings to modify prioritize potential mappings. Mappings are selected as fol- the target input script. The match rules are designed to allows: low only one difference in a match. The only difference between two nodes matched with the predicate-match rule is 1. The first pass through all pairs of mappings identifies one entity. Similarly, the only difference between two nodes iconic mappings. If the mapping involves an iconic node matched with the ordered-entity-match rule is the predicate. from the target, the mapping is discarded. The reason Consequently, this phase involves transfer of the difference is that iconicity is a distinguishing characteristic that we from the source to the target. The only noteworthy detail is want to preserve. For a clear illustration, a pirate is a that entity transfer is a global re-mapping of the entity akin character typically wearing a pirate’s hat, sword, and eye to a variable rebind. From figure 3 above, the predicate- patch. However, if you remove those accessories, the pi- match rule maps Driver (source script) → Passenger (target rate is no longer a pirate, but a relatively uninteresting script). It would be incoherent to map the entity only for that person. For the same reasons, iconicity in the source is a one specific node. Predicate changes, however, are local to characteristic that, if possible, we want to transfer to the the target node specified in the mapping. target. For this reason, if a mapping contains an iconic Figure 4 illustrates the output. All modifications to the source node, then we prioritize this mapping for selection. target input script are in bold. The blend script can be inter- 2. The second pass through the mappings sorts the mappings preted as a training day for a new tour bus driver.

47 begin( Cat begin( Lover, Day in Cowboy, 1 1 the life of a cat Cattle begin( Cat owner) Drive) Lover, 1 Cattle Drive) feed(Cat ride( ride( Cat Lover, 13 2 Cowboy, 2 Lover, 2 Cats) Horse) Horse) smack(0.9)( smack(0.9)( encounter( Cat slips( Cat Cowboy, slips( Cat Lover, slips( Cat 3 4 3 4 3 4 Lover, River) Lover) Rope, Cowboy) Rope, Lover) Saddle) Saddle) impact( frightened( impact( impact( 5 6 5 Cowboy, Stray Cat, River) Cat Lover, 5 Cat Lover, Ground) Ground) Ground) injures( wear( Cat Lover, injures( injures( 7 8 6 Cowboy, Gloves) Cat Lover, 6 Cat Lover, Arm) Arm) Arm) bandage( rescue( Fright- bandage( encounter( encounter( bandage( 9 10 8 7 Cowboy, ened Cat) Cat Lover, Cowboy, Cat Lover, 8 7 Cat Lover, Injury) River) Injury) River) Injury)

injures( Fright- frightened( frightened( ened Cat, Cat 11 9 Cow, Stray Cat, 9 Lover) River) River)

rescue( arrival( Cat rescue( 12 10 Frightened Lover, Home) Frightened 10 Cow) Cat) arrival( arrival( sneeze( Cat Cattle, Cattle, 13 11 11 Lover, Dander) Destina- Destina- tion) tion)

Figure 5: Cat owner and Cattle drive counterpart mappings Figure 6: “Cat Lover” + “Cattle Drive” output blend

4 Script Generation Examples

The first example, inspired by TV commercials, is a blend enter (Derrick enter( Bull, Sta- 1 1 between “A Day in the Life of a Cat Owner” (source input) Rose, Stadium) dium) and “Cattle Drive” (target input). Figure 5 depicts the source watch( Matador, rise(Crowd) 2 2 input (on the left), the target input (on the right), and selected Bull) counterpart mappings, all of which are created due to the predicate-match rule. Consequently, all of the changes to the drive( Derrick 3 3 applaud( Crowd) target script to produce the blended script are (global) entity Rose) substitutions. Figure 6 depicts the output, a script containing block( Opponent, charge( Bull, actions and events that could be summarized as “Cat Lover 4 4 Bull) Opponent) Joins the Cattle Drive”. All modifications to the original target input script are in bold. dodge( Derrick chase( Picadores, 5 5 The second example is a blend between a source script of Rose, Opponent) Bull) Derrick Rose, a member of the Chicago Bulls NBA basket- run( Bull, Oppo- intimidate( Op- ball team, charging the basketball hoop and a target script 6 6 of a bullfight. Figure 7 depicts the source and target in- nent) ponent, Bull) puts and selected counterpart mappings. A key aspects of juke( Bull, Op- raise( Matador, 7 7 this blend, which is not explicitly modeled computationally, ponent) Capote) but which is readily accessible to humans, is the linguistic double-meaning of the word “bull”, referring to both a bull drive( Target) 8 8 cheer( Crowd) in a bullfight and Derrick Rose as a Chicago Bull. Both meanings appear in the source and target scripts and allow dunk( Derrick 9 the algorithm to establish the “creative” counterpart map- Rose) pings depicted in Figure 7. Figure 8 illustrates the script blend, with predicate and entity substitutions in bold. cheer(Crowd) 10 5 Discussion Figure 7: Derrick Rose and Bullfight counterpart mappings In this section, we discuss a few noteworthy details and salient characteristics of our algorithm and approach.

48 enter ( Derrick 1 requirement for coherence of a blended script. Guaranteeing Rose, Stadium) causality, and thus coherence, requires a vast amount of general and speciﬁc knowledge that we do not currently use in watch( Matador, 2 Derrick Rose) our approach. As a result, our algorithm can produce incoherent blended scripts. Additionally, our algorithm produces

rise(Crowd) 3 blends only for two inputs scripts (and not n-ary inputs). Another limitation of our current approach is the use of a literal (syntactic) match when testing for equality between charge( Derrick 4 Rose, Opponent) the predicates or entities in two nodes. For a given knowledge base, this could be a signiﬁcant limitation to identify- chase(Picadores, ing existing counterpart mappings. This could, in theory, 5 Derrick Rose) be resolved in a reasonably straightforward manner by us-

block( Oppo- ing another equality test. For example, perhaps a minimum nent, Derrick 6 similarity threshold for two entities, given a (presumably) Rose) common prototype.

raise( Matador, 7 Capote) Future Work This paper presents first steps toward a script generation al- cheer(Crowd) 8 gorithm. We are currently undertaking an effort to acquire a script knowledge base in order to evaluate our approach. Figure 8: “Derrick Rose” + bullfight output blend Concurrent with that effort, we have identified a few possible extensions to our algorithm to ensure it is generally applicable. Our algorithm is currently designed to perform predicate First, our current algorithm uses only two operations to and entity substitutions. We chose these operations in or- produce a blend: 1) entity substitution and 2) predicate sub- der to produce coherent, executable blends. As mentioned stitution. These substitutions do not provide us with an abil- above, our algorithm is designed for use by an intelligent ity to augment (modify) the structure of the target graph. For agent in narrative contexts like pretend play. The algorithm example, it may make sense to “graft” nodes from the source assumes that the input scripts are both valid and executable, script into the target script in the case where there is an es- and it is vital that the algorithm produces executable scripts tablished counterpart mapping between nodes. This would as well. provide a significant additional capability of executing ac- A predicate substitution occurs when a mapping is created tions originally specified in the source script that have no as a result of an ordered-entity-match. In this case, all of the counterpart in the target. arguments match. Moreover, the predicate is valid in the Second, we intentionally designed the algorithm to pro- source script for the matched arguments. As a result, there duce one blended joint activity for execution by an intelli- is a “high” likelihood that the transfer of the predicate from gent agent. However, SME and other systems regularly pro- the source to the target will be both coherent and executable. duce multiple blends with a relevant notion of “score” that However, there is no guarantee, perhaps due to a mismatch could prove useful. As noted above, there is no guarantee of context or a lack of ability to perform the action in the that a given blend is executable. Thus, a score metric could blended script (perhaps due to missing pre-conditions for the be devised that measures executability using the play con- action to occur that existed in the source script, but do not text. exist in the target script). Third, our algorithm may benefit from using contextual Narratives are comprised of both causality and tempo- information, perhaps by employing a commonsense knowl- rality, which led us to consider how temporality and ac- edge base such as ConceptNet (Liu and Singh 2004). Con- tion/event ordering influences our perception of similarity sider a scenario where a “Cooking” source script contains between activities (Rimmon-Kenan 2002). Because our the predicate cook(Person, Food) and a “Picnic” target script scripts explicitly specify temporality, we ensure that coun- contains the predicate eat(Person, Food). A counterpart terpart mappings account for node ordering. This require- mapping can be established between the two nodes accord- ment, in turn, helps to ensure that the algorithm produces ing to the ordered-entity-match rule, resulting in a predicate coherent, executable scripts by avoiding the introduction of substitution in the “Picnic script” of “eat” → “cook”, which violations of pre-conditions that may result from reordered is incoherent because cooking cannot occur during a picnic, 2 nodes. due to a lack of a heat source . A similar problem occurs for entity substitution. An additional post-processing step Limitations could be used to identify, and subsequently filter, “inappro- priate” blends based on context, which could be knowledge As discussed above, our algorithm uses two features of intensive. scripts, temporality and iconicity, during blending. These surface-level features do not provide us with the semantic 2We admit that there could be exceptions, such as a picnic with information required to maintain causality, which is a strong a fire, but a canonical picnic does not involve a fire.

49 References Falkenhainer, B.; Forbus, K. D.; and Gentner, D. 1989. The structure-mapping engine: Algorithm and examples. Artifi- cial intelligence 41(1):1–63. Fauconnier, G., and Turner, M. 1998. Conceptual integration networks. Cognitive science 22(2):133–187. Fauconnier, G., and Turner, M. 2003. The way we think: Conceptual blending and the mind’s hidden complexities. Basic Books (AZ). Forbus, K. D.; Gentner, D.; and Law, K. 1995. MAC/FAC: a model of similarity-based retrieval. Cognitive Science 19(2):141–205. Gentner, D. 1983. Structure-mapping: A theoretical frame- work for analogy. Cognitive science 7(2):155–70. Goguen, J., and Harrell, F. 2004. Foundations for active multimedia narrative: Semiotic spaces and structural blending. Interaction Studies: Social Behaviour and Communi- cation in Biological and Artificial Systems. Harrell, D. F. 2005. Shades of computational evocation and meaning: The GRIOT system and improvisational poetry generation. In Proceedings, Sixth Digital Arts and Culture Conference, 133–143. Holyoak, K. J., and Thagard, P. 1989. Analogical mapping by constraint satisfaction. Cognitive science 13(3):295–355. Li, B.; Zook, A.; Davis, N.; and Riedl, M. O. 2012. Goal- driven conceptual blending: A computational approach for creativity. In International Conference on Computational Creativity, 10. Liu, H., and Singh, P. 2004. ConceptNet – a practical commonsense reasoning tool-kit. BT Technology Journal 22(4):211–226. Magerko, B.; Manzoul, W.; Riedl, M.; Baumer, A.; Fuller, D.; Luther, K.; and Pearce, C. 2009. An empirical study of cognition and theatrical improvisation. In Proceedings of the seventh ACM conference on Creativity and cognition, 117–126. Magerko, B.; Dohogne, P.; and DeLeon, C. 2011. Employ- ing fuzzy concepts for digital improvisational theatre. In Proceedings of the Seventh Annual International Artificial Intelligence and Interactive Digital Entertainment Confer- ence. Moran, S.; John-Steiner, V.; and Sawyer, R. K. 2003. Cre- ativity in the making. Creativity and development 61–90. Rimmon-Kenan, S. 2002. Narrative fiction: Contemporary poetics (new accents). Schank, R. C., and Abelson, R. P. 1977. Scripts, plans, goals and understanding: An inquiry into human knowledge structures. Veale, T.; O’Donoghue, D.; and Keane, M. T. 2000. Compu- tation and blending. Cognitive Linguistics 11(3/4):253–282. Zook, A.; Magerko, B.; and Riedl, M. 2011. Formally modeling pretend object play. In Proceedings of the 8th ACM conference on Creativity and cognition, C&C ’11, 147–156. New York, NY, USA: ACM.